Predicting Suicidal Ideation in Postpartum Women Using Perceived Stress and Social Support: A Machine Learning Approach

Objectives: To evaluate perceived stress and multidimensional social support as predictors of suicidal ideation in postpartum women, and to develop an explainable, uncertainty-aware machine learning framework for individualized risk screening. Method: We analyzed data from 200 postpartum women, including perceived stress (Perceived Stress Scale, 10 items), multidimensional social support (emotional, positive interaction, affective, and instrumental subscales), and demographic and clinical variables. We conducted descriptive comparisons, univariate and multivariate logistic regression, and a machine learning pipeline comparing seven classifiers under stratified five-fold cross-validation with Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance (17% prevalence of suicidal ideation). The final Random Forest and Logistic Regression models were evaluated on a held-out test set, calibrated, and interpreted using SHAP and LIME. Split conformal prediction generated statistically valid, uncertainty-aware prediction sets, and counterfactual sensitivity analysis quantified actionable score changes associated with reduced predicted risk. Findings: Women reporting suicidal ideation had significantly higher perceived stress (mean 36.6 vs. 29.6, p<0.001) and lower total social support (mean 49.8 vs. 65.4, p<0.001) than women without ideation. In multivariate analysis, perceived stress remained independently associated with suicidal ideation (adjusted odds ratio 1.23, 95% CI 1.11-1.37, p<0.001). The Random Forest classifier achieved the best cross-validated discrimination (ROC-AUC 0.766) and test-set performance (ROC-AUC 0.804, accuracy 0.84). SHAP analysis identified perceived stress and social support subscales as the dominant predictors of risk. Conformal prediction achieved 98% empirical coverage at a 90% target, with 34% of test cases flagged as ambiguous rather than confidently classified. Novelty: To our knowledge, this is the first postpartum suicidal ideation risk model to combine SHAP and LIME explainability with split conformal prediction and counterfactual sensitivity analysis in a single framework, jointly quantifying which factors drive each prediction, how confident the model is in that prediction, and what modifiable change would most plausibly lower a given patient's predicted risk. Keywords: postpartum depression, suicidal ideation, perceived stress, social support, machine learning, explainable AI, conformal prediction